@MinLiBuilds: This week, I spent 80 GPU hours running three rounds of tests on Qwen 3.8 Flash Next 125B. I tested 7 quantized weights…
Summary
@MinLiBuilds 花费 80 GPU 小时对 Qwen 3.8 Flash Next 125B 进行三轮测试，评估 7 种量化权重并给出针对不同显卡的部署建议，同时介绍能将 125B 模型塞入消费级显存的 Strata 推理框架。
View Cached Full Text
Cached at: 10/10/26, 10:47 AM
This week, I spent 80 GPU hours running three rounds of tests on Qwen 3.8 Flash Next 125B.
I tested 7 quantized weights and provided corresponding deployment recommendations: 4090 48G: UD-IQ4_XS / UD-Q4_K_XL
5090 5080: GSQ-RCO IQ3_S
8G/12G/16G cards: ISTA IQ2_XS > ISTA Q2_0 > Swift 1.5 IQ2_XS > Swift 1.5 Q2_0
This is a super interesting model—powerful and massive.
It’s even stronger than the hottest Qwen 3.8 27B -Dense.
🔥 The new inference framework Strata gets every cell working—GPU, CPU, memory, disk, motherboard bus—managing to cram the impossible-to-fit 125B model into consumer-grade VRAM.
Detailed test post is in the thread below 👇
King of Quality: UD-Q4_K_XL UD-IQ4_XS GSQ-RCO IQ3_S
速度强者： UD-IQ4_XS UD-Q4_K_XL
2bit最佳选择：
Swift 1.5 Q2_0 Swift 1.5 IQ2_XS ISTA Q2_0 ISTA IQ2_XS
你对并发有要求么。 是时候加一些内存了。
我推荐iq4，不过你的内存在临界点，可以跑GSQ-RCO IQ3_S
太好了，有更新我会跟进。
ISTA IQ2_XS可以跑，编程25/30，iQ4是28/30 具体可以使用256K上下文，实测一下速度
With such a large amount of memory, you could try testing the speed. I’d suggest iQ4 and GSQ-RCO IQ3_S—these two are worth trying out.
It’d be great if GSQ-RCO had iQ4_S.
内存有点少，可以跟进一下最新的，能不能有二级缓存
实测一下，关键是ctx长度，速度应该不会太慢
Similar Articles
@MinLiBuilds: https://x.com/MinLiBuilds/status/2107988055101685781
The author shares a complete guide for deploying Qwen3.8-Flash-Next (125B-A6B) on 12GB VRAM (RTX 5070) with Strata, achieving nearly 100 tok/s, and lets Ant Group's newly released Ling-3.1-flash Agent handle the entire workflow from scratch — machine check, download, installation, troubleshooting, and benchmarking.
@MinLiBuilds: The GGUF version of Qwen3.8-Flash-Next, an 111GB model, surprisingly runs on dual RTX 4090 with 96GB VRAM. This model is truly bizarre. My setup: 2× RTX 4090 48G = 96GB VRAM, 62GB RAM, using Unsloth's …
Successfully ran the 111GB Qwen3.8-Flash-Next GGUF model on dual RTX 4090 with 96GB VRAM, achieving a generation speed of 66 tokens/s through PLE and sparse attention optimizations, with detailed explanations of quantization and runtime configuration.
Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s
Strata 是一个免费开源的推理引擎，可在 RTX 4090 等消费级显卡（12GB+ 显存）本地运行 Qwen3.8-Flash-Next（125B）模型，写入速度约 100 tokens/s，且支持 Windows/Linux 和 NVIDIA/AMD 显卡。
@MinLiBuilds: I just finished running a round of experiments with Strata. This is truly an awesome inference framework—VRAM is still …
A creator shares benchmark experiments with the Strata inference framework, systematically limiting VRAM budgets for the expert cache on an RTX 4090 to measure how speed degrades across different simulated GPU memory tiers, using Qwen3.8-Flash-Next Q2_0 at 8192 context and temperature 0.
@MinLiBuilds: https://x.com/MinLiBuilds/status/2089338660386992295
This article compares the performance of NVIDIA DGX Spark and a modified RTX 4090 in locally deploying the Qwen3.8-27B and Ling-3.0-flash models, providing benchmark data and purchase recommendations.